A Transformer-Based Versatile Network for Acne Vulgaris Segmentation

One of the most typical skin disorders is acne. Finding acne problems to treat is a complicated process. Dermatologists use manual skin evaluation techniques, such as visual and photography inspection, to find it. Acne on the patient's face is manually marked and counted, a laborious and arbitrary process. Several methods for spotting acne have been developed recently, including various machine learning and image processing methods. These methods start with image capture and acne segmentation. This paper introduces a novel versatile transformer-based automated segmentation model that segments acne diseases and classifies acne types. This method consists of a dual encoder, a feature versatile block (FVB), and efficient decoder architecture with skip connections. The dual encoder combines transformer and CNN, extracting and encoding rich local characteristics while simultaneously capturing crucial global context data for segmenting acne lesions. Then, the FVB is applied to distribute and integrate retrieved features between the encoder and decoder, which are also adaptively matched using skip connection. Finally, an efficient decoder is proposed to segment and classify acne diseases. Experimental results show that our method outperforms other deep learning-based segmentation methods.